以ALS(Airborne Laser Scanning)系统为代表的三维点云获取技术为建筑物重建提供了一条高效便捷途径。本文基于ALS点云超体素，提出了一种方差分析引导的高阶马尔可夫网络，并用以提取建筑物。该方法以超体素作为无向图模型节点，根据三维邻域特有的局部几何属性，结合方差分析原理，生成高阶因子，并将特征转化为表达能力更强的节点和边势函数；再采用信念传播算法，对高阶马尔可夫网络进行推理，形成了一种非监督的建筑物识别框架。此外，本文采用由粗到精的识别策略，首先在独立假设前提下，利用贝叶斯高斯混合模型实现节点初始状态捕捉，再采用高阶马尔可夫网络对三维邻域的相关性建模，以提高建筑物提取精度。本文引入两组具备人工标记真值的开放ALS点云数据集进行实验，利用四种国际通用指标对建筑物提取结果进行精度评价。可视化分析表明，本文方法提取的建筑物内部完整，且边界清晰，为建筑物三维重建提供了可靠信息。定量化分析表明，在低层建筑主导的住宅区,本文方法的建筑物提取精度（基于投影面积和对象的F1指数）为95.4%和91.5%，均高于现有监督和非监督识别方法；在高层建筑主导的商业区，本文方法基于对象的F1指数取得93.5%，高于现有方法，基于投影面积的F1指数为92.9%，仍处于较高水平。

Three-dimensional (3D) point cloud acquisition technology represented by Airborne Laser Scanning (ALS) system provides an efficient and convenient way for reconstruction of buildings. In this paper, based on the supervoxels of ALS point clouds, a high-order Markov network guided by the analysis of variance (ANOVA) is proposed to extract buildings. This method first uses supervoxels as the nodes of the undirected graph, and then constructs high-order factors based on the principle of ANOVA and the local geometric features of three-dimensional neighborhood. After that, the features are transformed into node and edge potential functions with a better expression ability. Finally, the belief propagation algorithm is used to make inference on the high-order Markov network, and an unsupervised building recognition framework is formed. In addition, this unsupervised approach recognizes buildings from coarse to fine. Specifically, the Bayesian Gaussian Mixture Model is first used to capture the initial state of the node under the independent assumption, and then the high-order Markov network is designed to model the correlation of the 3D neighborhood for extracting buildings. In the experiments, two groups of ALS point cloud datasets with ground-truths are employed, and four commonly used metrics are utilized to evaluate the accuracy of the results. Visual analysis shows that our method extracts buildings with complete interiors and clear boundaries. Hence, this method can be used to provide reliable data for the 3D reconstruction of buildings. According to quantitative analysis, in residential areas dominated by low-rise buildings, the averaged accuracy (the projection-area-based and object-based F1 indexes) of our method reaches 95.4% and 91.5% which are higher than that of existing supervised and unsupervised methods. In downtown areas dominated by high-rise buildings, the averaged object-based F1 score of our method reaches 93.5% which is higher than that of existing methods; and its averaged projection-area-based F1 score gets 92.9%, which is more than sufficient.